diff --git a/deepspeech/exps/deepspeech2/bin/train.py b/deepspeech/exps/deepspeech2/bin/train.py
index 69ff043a08d28171711543afcbd51bcd571e69d2..6740f288fd667d2f84bca96277c46da9ce894c42 100644
--- a/deepspeech/exps/deepspeech2/bin/train.py
+++ b/deepspeech/exps/deepspeech2/bin/train.py
@@ -27,7 +27,7 @@ def main_sp(config, args):
 
 
 def main(config, args):
-    if args.device == "gpu" and args.nprocs > 1:
+    if args.nprocs > 0:
         dist.spawn(main_sp, args=(config, args), nprocs=args.nprocs)
     else:
         main_sp(config, args)
diff --git a/deepspeech/exps/deepspeech2/model.py b/deepspeech/exps/deepspeech2/model.py
index 7bf029300ff16e56a08aee8d73c22ae9c0468479..8af2b02ab0856947d35650a9f6ae42d0b532885d 100644
--- a/deepspeech/exps/deepspeech2/model.py
+++ b/deepspeech/exps/deepspeech2/model.py
@@ -403,7 +403,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
     def setup(self):
         """Setup the experiment.
         """
-        paddle.set_device(self.args.device)
+        paddle.set_device('gpu' self.args.nprocs > 0 else 'cpu')
 
         self.setup_output_dir()
         self.setup_checkpointer()
@@ -635,7 +635,7 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
     def setup(self):
         """Setup the experiment.
         """
-        paddle.set_device(self.args.device)
+        paddle.set_device('gpu' self.args.nprocs > 0 else 'cpu')
 
         self.setup_output_dir()
 
diff --git a/deepspeech/exps/u2/bin/train.py b/deepspeech/exps/u2/bin/train.py
index b664401a2f21e7c12931868cd10bdefbb536c9f3..17fb08a6c41202da97fec48911ba714942cf29c7 100644
--- a/deepspeech/exps/u2/bin/train.py
+++ b/deepspeech/exps/u2/bin/train.py
@@ -32,7 +32,7 @@ def main_sp(config, args):
 
 
 def main(config, args):
-    if args.device == "gpu" and args.nprocs > 1:
+    if args.nprocs > 0:
         dist.spawn(main_sp, args=(config, args), nprocs=args.nprocs)
     else:
         main_sp(config, args)
diff --git a/deepspeech/exps/u2/model.py b/deepspeech/exps/u2/model.py
index 6bf0190050cbe6ec16ba6d7983146e8b62893713..bc46a104b6b6b73689378b473a05698718e8d3b2 100644
--- a/deepspeech/exps/u2/model.py
+++ b/deepspeech/exps/u2/model.py
@@ -653,7 +653,7 @@ class U2Tester(U2Trainer):
     def setup(self):
         """Setup the experiment.
         """
-        paddle.set_device(self.args.device)
+        paddle.set_device('gpu' self.args.nprocs > 0 else 'cpu')
 
         self.setup_output_dir()
         self.setup_checkpointer()
diff --git a/deepspeech/exps/u2_kaldi/bin/train.py b/deepspeech/exps/u2_kaldi/bin/train.py
index 1dcd154d35bcd941db66260fe54e60873694dc28..d909727f3a5a637f977b2f5569d553db5ff382d1 100644
--- a/deepspeech/exps/u2_kaldi/bin/train.py
+++ b/deepspeech/exps/u2_kaldi/bin/train.py
@@ -36,7 +36,7 @@ def main_sp(config, args):
 
 
 def main(config, args):
-    if args.device == "gpu" and args.nprocs > 1:
+    if args.nprocs > 0:
         dist.spawn(main_sp, args=(config, args), nprocs=args.nprocs)
     else:
         main_sp(config, args)
diff --git a/deepspeech/exps/u2_kaldi/model.py b/deepspeech/exps/u2_kaldi/model.py
index bc7cd4fd3671012cae3eb3d5913188214dcba8e0..be89c3d694255e68a8893e6de77b352d925cdc0a 100644
--- a/deepspeech/exps/u2_kaldi/model.py
+++ b/deepspeech/exps/u2_kaldi/model.py
@@ -637,7 +637,7 @@ class U2Tester(U2Trainer):
     def setup(self):
         """Setup the experiment.
         """
-        paddle.set_device(self.args.device)
+        paddle.set_device('gpu' self.args.nprocs > 0 else 'cpu')
 
         self.setup_output_dir()
         self.setup_checkpointer()
diff --git a/deepspeech/exps/u2_st/bin/train.py b/deepspeech/exps/u2_st/bin/train.py
index 86a0f000051c720276fccdadcf5d8cd0e27a9c9c..1e6a746b848ed7e387aec91217c96640e628a39e 100644
--- a/deepspeech/exps/u2_st/bin/train.py
+++ b/deepspeech/exps/u2_st/bin/train.py
@@ -30,7 +30,7 @@ def main_sp(config, args):
 
 
 def main(config, args):
-    if args.device == "gpu" and args.nprocs > 1:
+    if args.nprocs > 0:
         dist.spawn(main_sp, args=(config, args), nprocs=args.nprocs)
     else:
         main_sp(config, args)
diff --git a/deepspeech/exps/u2_st/model.py b/deepspeech/exps/u2_st/model.py
index 4f95bc42b35491329f183c3e4cd3065ba29212de..55dadee8229015b27a674411a71adb8809c2f507 100644
--- a/deepspeech/exps/u2_st/model.py
+++ b/deepspeech/exps/u2_st/model.py
@@ -661,7 +661,7 @@ class U2STTester(U2STTrainer):
     def setup(self):
         """Setup the experiment.
         """
-        paddle.set_device(self.args.device)
+        paddle.set_device('gpu' self.args.nprocs > 0 else 'cpu')
 
         self.setup_output_dir()
         self.setup_checkpointer()
diff --git a/deepspeech/training/cli.py b/deepspeech/training/cli.py
index 07c213dbca1e0ecb5ba0bf37b71df0859f35ceb6..aa263a0666a611c315803c1782bb45fe4ae5d70a 100644
--- a/deepspeech/training/cli.py
+++ b/deepspeech/training/cli.py
@@ -30,7 +30,7 @@ def default_argument_parser():
 
     The ``--checkpoint_path`` specifies the checkpoint to load from.
 
-    The ``--device`` and ``--nprocs`` specifies how to run the training.
+    The ``--nprocs`` specifies how to run the training.
 
 
     See Also
@@ -51,12 +51,6 @@ def default_argument_parser():
         default=None,
         help="seed to use for paddle, np and random. None or 0 for random, else set seed."
     )
-    train_group.add_argument(
-        "--device",
-        type=str,
-        default='gpu',
-        choices=["cpu", "gpu"],
-        help="device cpu and gpu are supported.")
     train_group.add_argument(
         "--nprocs",
         type=int,
diff --git a/deepspeech/training/trainer.py b/deepspeech/training/trainer.py
index 8b1adcd037cb39eae47ad115d59c50e41fc357f0..b180f4890feb60d083291febc57e1456792802b9 100644
--- a/deepspeech/training/trainer.py
+++ b/deepspeech/training/trainer.py
@@ -86,7 +86,7 @@ class Trainer():
     >>>     config.merge_from_list(args.opts)
     >>> config.freeze()
     >>>
-    >>> if args.nprocs > 1 and args.device == "gpu":
+    >>> if args.nprocs > 0:
     >>>     dist.spawn(main_sp, args=(config, args), nprocs=args.nprocs)
     >>> else:
     >>>     main_sp(config, args)
@@ -119,7 +119,7 @@ class Trainer():
     def setup(self):
         """Setup the experiment.
         """
-        paddle.set_device(self.args.device)
+        paddle.set_device('gpu' self.args.nprocs > 0 else 'cpu')
         if self.parallel:
             self.init_parallel()
 
@@ -139,7 +139,7 @@ class Trainer():
         """A flag indicating whether the experiment should run with
         multiprocessing.
         """
-        return self.args.device == "gpu" and self.args.nprocs > 1
+        return elf.args.nprocs > 0
 
     def init_parallel(self):
         """Init environment for multiprocess training.
diff --git a/examples/aishell/s0/local/export.sh b/examples/aishell/s0/local/export.sh
index 2e09e5f5e76a7f7cdf9cca8fbb91d66bb48aea0c..a5e62c28d2fa23a5ef9b9e2a0281b025aa943a30 100755
--- a/examples/aishell/s0/local/export.sh
+++ b/examples/aishell/s0/local/export.sh
@@ -13,13 +13,7 @@ ckpt_path_prefix=$2
 jit_model_export_path=$3
 model_type=$4
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/export.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/aishell/s0/local/test.sh b/examples/aishell/s0/local/test.sh
index 9fd0bc8d5bcdded1d33990a8ae20101d0a538441..2ae0740b3e8d44ab03e45f4c1b5dbb945657705e 100755
--- a/examples/aishell/s0/local/test.sh
+++ b/examples/aishell/s0/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 model_type=$3
@@ -23,8 +19,7 @@ if [ $? -ne 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/test.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/aishell/s0/local/test_export.sh b/examples/aishell/s0/local/test_export.sh
index b6d580979c9797d7de59b7555c7304e00658b591..a9a6b122df8055f872f9f0a68717b57241d99359 100755
--- a/examples/aishell/s0/local/test_export.sh
+++ b/examples/aishell/s0/local/test_export.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 jit_model_export_path=$2
 model_type=$3
@@ -23,8 +19,7 @@ if [ $? -ne 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/test_export.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${jit_model_export_path}.rsl \
 --export_path ${jit_model_export_path} \
diff --git a/examples/aishell/s0/local/train.sh b/examples/aishell/s0/local/train.sh
index 668ad0ead41bc9933e2fd63802b95ec5263999d2..edbf3383070ed0c8a71a125c0e77371d0a4412b6 100755
--- a/examples/aishell/s0/local/train.sh
+++ b/examples/aishell/s0/local/train.sh
@@ -12,11 +12,6 @@ config_path=$1
 ckpt_name=$2
 model_type=$3
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 mkdir -p exp
 
 # seed may break model convergence
@@ -26,7 +21,6 @@ if [ ${seed} != 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/train.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/aishell/s1/local/align.sh b/examples/aishell/s1/local/align.sh
index ad6c84bc8cf398cddecfadbbb47b7ef9c60e9158..279461aafe19967ac054df815b08e60e351fcc7f 100755
--- a/examples/aishell/s1/local/align.sh
+++ b/examples/aishell/s1/local/align.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 
@@ -22,8 +18,7 @@ mkdir -p ${output_dir}
 # align dump in `result_file`
 # .tier, .TextGrid dump in `dir of result_file`
 python3 -u ${BIN_DIR}/alignment.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/aishell/s1/local/export.sh b/examples/aishell/s1/local/export.sh
index f99a15bade1c89f968e84a6c10d500466f884d5b..b562218e7a76c91d3b906d2c099218bf492627a8 100755
--- a/examples/aishell/s1/local/export.sh
+++ b/examples/aishell/s1/local/export.sh
@@ -12,13 +12,7 @@ config_path=$1
 ckpt_path_prefix=$2
 jit_model_export_path=$3
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/export.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/aishell/s1/local/test.sh b/examples/aishell/s1/local/test.sh
index f7e99ad7f668d804a108e295faf9ac2e58760595..c87412c9b67533f1ff3275097099f628f964a608 100755
--- a/examples/aishell/s1/local/test.sh
+++ b/examples/aishell/s1/local/test.sh
@@ -8,11 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 config_path=$1
 ckpt_prefix=$2
 
@@ -39,8 +34,7 @@ for type in attention ctc_greedy_search; do
     output_dir=${ckpt_prefix}
     mkdir -p ${output_dir}
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
@@ -58,8 +52,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     output_dir=${ckpt_prefix}
     mkdir -p ${output_dir}
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/aishell/s1/local/train.sh b/examples/aishell/s1/local/train.sh
index 5097d4d030cab74fe5c3d52b223eddd1c27ea68d..71af3a006deec9032fcefd561fc46cea274d1e10 100755
--- a/examples/aishell/s1/local/train.sh
+++ b/examples/aishell/s1/local/train.sh
@@ -12,11 +12,6 @@ source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 if [ ${seed} != 0  ]; then
     export FLAGS_cudnn_deterministic=True
     echo "using seed $seed & FLAGS_cudnn_deterministic=True ..."
@@ -34,7 +29,6 @@ mkdir -p exp
 
 python3 -u ${BIN_DIR}/train.py \
 --seed ${seed} \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/callcenter/s1/local/align.sh b/examples/callcenter/s1/local/align.sh
index f2c878c20c6d5bb9a89b437c6c0d5b11b1b80c4f..b679e2ea7fcb1d5ebd5a50dc88cb90e59e3e4789 100755
--- a/examples/callcenter/s1/local/align.sh
+++ b/examples/callcenter/s1/local/align.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 
@@ -20,7 +16,6 @@ ckpt_name=$(basename ${ckpt_prefxi})
 mkdir -p exp
 
 
-
 batch_size=1
 output_dir=${ckpt_prefix}
 mkdir -p ${output_dir}
@@ -28,8 +23,7 @@ mkdir -p ${output_dir}
 # align dump in `result_file`
 # .tier, .TextGrid dump in `dir of result_file`
 python3 -u ${BIN_DIR}/alignment.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/callcenter/s1/local/export.sh b/examples/callcenter/s1/local/export.sh
index d171899cdbf4220436bb71ad07e3a1a5e9ea8bc2..d5f912e9033b1d71423d87ef0a9cf4b1991d3563 100755
--- a/examples/callcenter/s1/local/export.sh
+++ b/examples/callcenter/s1/local/export.sh
@@ -12,13 +12,7 @@ config_path=$1
 ckpt_path_prefix=$2
 jit_model_export_path=$3
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/export.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/callcenter/s1/local/test.sh b/examples/callcenter/s1/local/test.sh
index 7a5b1cdb1c6f62fe25e6349226198ea6d8868a5e..dca3137dd0cf1faadd6b3c383dcb178881bb9080 100755
--- a/examples/callcenter/s1/local/test.sh
+++ b/examples/callcenter/s1/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 
@@ -32,8 +28,7 @@ for type in attention ctc_greedy_search; do
     output_dir=${ckpt_prefix}
     mkdir -p ${output_dir}
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
@@ -51,8 +46,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     output_dir=${ckpt_prefix}
     mkdir -p ${output_dir}
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/callcenter/s1/local/train.sh b/examples/callcenter/s1/local/train.sh
index d5dc15b03a32c2c01ba8850b38dbb2c4816cc80c..eb8f86626f8988890c89dc29accb5ae1f172e1d6 100755
--- a/examples/callcenter/s1/local/train.sh
+++ b/examples/callcenter/s1/local/train.sh
@@ -11,10 +11,6 @@ echo "using $ngpu gpus..."
 config_path=$1
 ckpt_name=$2
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 echo "using ${device}..."
 
 mkdir -p exp
@@ -26,7 +22,6 @@ if [ ${seed} != 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/train.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/librispeech/s0/local/export.sh b/examples/librispeech/s0/local/export.sh
index 2e09e5f5e76a7f7cdf9cca8fbb91d66bb48aea0c..a5e62c28d2fa23a5ef9b9e2a0281b025aa943a30 100755
--- a/examples/librispeech/s0/local/export.sh
+++ b/examples/librispeech/s0/local/export.sh
@@ -13,13 +13,7 @@ ckpt_path_prefix=$2
 jit_model_export_path=$3
 model_type=$4
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/export.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/librispeech/s0/local/test.sh b/examples/librispeech/s0/local/test.sh
index b5b68c599c45ab50aa12ee35c120e02fb68740b4..4d00f30b852da5a370f5d4934f3caadd2b833c00 100755
--- a/examples/librispeech/s0/local/test.sh
+++ b/examples/librispeech/s0/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 model_type=$3
@@ -23,8 +19,7 @@ if [ $? -ne 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/test.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/librispeech/s0/local/train.sh b/examples/librispeech/s0/local/train.sh
index 6aee372a423aa7c037eb8def8c60086606d9a614..519df7fe9387a215f6d8835d90869ec0d4d7f5e5 100755
--- a/examples/librispeech/s0/local/train.sh
+++ b/examples/librispeech/s0/local/train.sh
@@ -12,12 +12,6 @@ config_path=$1
 ckpt_name=$2
 model_type=$3
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-echo "using ${device}..."
-
 mkdir -p exp
 
 # seed may break model convergence
@@ -27,7 +21,6 @@ if [ ${seed} != 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/train.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/librispeech/s1/local/align.sh b/examples/librispeech/s1/local/align.sh
index ad6c84bc8cf398cddecfadbbb47b7ef9c60e9158..279461aafe19967ac054df815b08e60e351fcc7f 100755
--- a/examples/librispeech/s1/local/align.sh
+++ b/examples/librispeech/s1/local/align.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 
@@ -22,8 +18,7 @@ mkdir -p ${output_dir}
 # align dump in `result_file`
 # .tier, .TextGrid dump in `dir of result_file`
 python3 -u ${BIN_DIR}/alignment.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/librispeech/s1/local/export.sh b/examples/librispeech/s1/local/export.sh
index f99a15bade1c89f968e84a6c10d500466f884d5b..b562218e7a76c91d3b906d2c099218bf492627a8 100755
--- a/examples/librispeech/s1/local/export.sh
+++ b/examples/librispeech/s1/local/export.sh
@@ -12,13 +12,7 @@ config_path=$1
 ckpt_path_prefix=$2
 jit_model_export_path=$3
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/export.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/librispeech/s1/local/test.sh b/examples/librispeech/s1/local/test.sh
index 3bd3f0bba7668b4b52be17b64bbf415d6dc627a3..f7ec34ab57642b456247279984a1c761263d773a 100755
--- a/examples/librispeech/s1/local/test.sh
+++ b/examples/librispeech/s1/local/test.sh
@@ -8,11 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 config_path=$1
 ckpt_prefix=$2
 
@@ -38,8 +33,7 @@ for type in attention ctc_greedy_search; do
         batch_size=64
     fi
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
@@ -55,8 +49,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     echo "decoding ${type}"
     batch_size=1
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/librispeech/s1/local/train.sh b/examples/librispeech/s1/local/train.sh
index f905b766ef52eb1ffcd98b9f31109ace916a5889..8f92c64694662b99a681bbf709f8da60d4e0eb0d 100755
--- a/examples/librispeech/s1/local/train.sh
+++ b/examples/librispeech/s1/local/train.sh
@@ -11,12 +11,6 @@ echo "using $ngpu gpus..."
 config_path=$1
 ckpt_name=$2
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-echo "using ${device}..."
-
 mkdir -p exp
 
 # seed may break model convergence
@@ -25,8 +19,10 @@ if [ ${seed} != 0 ]; then
     export FLAGS_cudnn_deterministic=True
 fi
 
+# export FLAGS_cudnn_exhaustive_search=true
+# export FLAGS_conv_workspace_size_limit=4000
+
 python3 -u ${BIN_DIR}/train.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/librispeech/s2/local/align.sh b/examples/librispeech/s2/local/align.sh
index b3d8fa5f5519a00cf22b1940ab2eec110ebfb21e..b45f4a0f5c270740189b3a917b335eca7efdd164 100755
--- a/examples/librispeech/s2/local/align.sh
+++ b/examples/librispeech/s2/local/align.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 dict_path=$2
 ckpt_prefix=$3
@@ -26,8 +22,7 @@ python3 -u ${BIN_DIR}/test.py \
 --model-name 'u2_kaldi' \
 --run-mode 'align' \
 --dict-path ${dict_path} \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result-file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/librispeech/s2/local/export.sh b/examples/librispeech/s2/local/export.sh
index efa70a2b961b0a7b8e05b3008c82fea5ec94872d..9c66dc62aa603aee0df2d48e90c8c7fe955063c9 100755
--- a/examples/librispeech/s2/local/export.sh
+++ b/examples/librispeech/s2/local/export.sh
@@ -12,15 +12,9 @@ config_path=$1
 ckpt_path_prefix=$2
 jit_model_export_path=$3
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/test.py \
 --model-name 'u2_kaldi' \
 --run-mode 'export' \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/librispeech/s2/local/train.sh b/examples/librispeech/s2/local/train.sh
index 66754201f7bfbf6f0b99cdd13ca20617b4b1d8bd..33b46c20f1f752c158fb304ce8470e88d6d4a761 100755
--- a/examples/librispeech/s2/local/train.sh
+++ b/examples/librispeech/s2/local/train.sh
@@ -11,12 +11,6 @@ echo "using $ngpu gpus..."
 config_path=$1
 ckpt_name=$2
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-echo "using ${device}..."
-
 mkdir -p exp
 
 # seed may break model convergence
@@ -27,7 +21,6 @@ fi
 
 python3 -u ${BIN_DIR}/train.py \
 --model-name u2_kaldi \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/ted_en_zh/t0/local/test.sh b/examples/ted_en_zh/t0/local/test.sh
index 642328e88b221e6bb2b322a0696884bd478d27ff..34475085538b2b7994d9955003aed2704fbb31ca 100755
--- a/examples/ted_en_zh/t0/local/test.sh
+++ b/examples/ted_en_zh/t0/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 
@@ -19,8 +15,7 @@ for type in fullsentence; do
     echo "decoding ${type}"
     batch_size=32
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/ted_en_zh/t0/local/train.sh b/examples/ted_en_zh/t0/local/train.sh
index f905b766ef52eb1ffcd98b9f31109ace916a5889..e5fd19ddb8d523acc328f13e3a05c126ae56db29 100755
--- a/examples/ted_en_zh/t0/local/train.sh
+++ b/examples/ted_en_zh/t0/local/train.sh
@@ -11,12 +11,6 @@ echo "using $ngpu gpus..."
 config_path=$1
 ckpt_name=$2
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-echo "using ${device}..."
-
 mkdir -p exp
 
 # seed may break model convergence
@@ -26,7 +20,6 @@ if [ ${seed} != 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/train.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/timit/s1/local/align.sh b/examples/timit/s1/local/align.sh
index ad6c84bc8cf398cddecfadbbb47b7ef9c60e9158..279461aafe19967ac054df815b08e60e351fcc7f 100755
--- a/examples/timit/s1/local/align.sh
+++ b/examples/timit/s1/local/align.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 
@@ -22,8 +18,7 @@ mkdir -p ${output_dir}
 # align dump in `result_file`
 # .tier, .TextGrid dump in `dir of result_file`
 python3 -u ${BIN_DIR}/alignment.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/timit/s1/local/export.sh b/examples/timit/s1/local/export.sh
index f99a15bade1c89f968e84a6c10d500466f884d5b..b562218e7a76c91d3b906d2c099218bf492627a8 100755
--- a/examples/timit/s1/local/export.sh
+++ b/examples/timit/s1/local/export.sh
@@ -12,13 +12,7 @@ config_path=$1
 ckpt_path_prefix=$2
 jit_model_export_path=$3
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/export.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/timit/s1/local/test.sh b/examples/timit/s1/local/test.sh
index a137924e22ac10328fb79e08536f65a9f2379418..868c8fda8073a8bb5bca49bf57930cb9274c5a1b 100755
--- a/examples/timit/s1/local/test.sh
+++ b/examples/timit/s1/local/test.sh
@@ -8,11 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 config_path=$1
 ckpt_prefix=$2
 
@@ -37,8 +32,7 @@ for type in attention ctc_greedy_search; do
         batch_size=64
     fi
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
@@ -54,8 +48,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     echo "decoding ${type}"
     batch_size=1
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu}  \
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/timit/s1/local/train.sh b/examples/timit/s1/local/train.sh
index 180d8b5a7e4e38981fb3c087c6c754ebdfe14b71..89a64327c1befc875f50fed37f93b4abd021ac8e 100755
--- a/examples/timit/s1/local/train.sh
+++ b/examples/timit/s1/local/train.sh
@@ -11,12 +11,6 @@ echo "using $ngpu gpus..."
 config_path=$1
 ckpt_name=$2
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-echo "using ${device}..."
-
 mkdir -p exp
 
 # seed may break model convergence
@@ -26,7 +20,6 @@ if [ ${seed} != 0  ]; then
 fi
 
 python3 -u ${BIN_DIR}/train.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/tiny/s0/local/export.sh b/examples/tiny/s0/local/export.sh
index 2e09e5f5e76a7f7cdf9cca8fbb91d66bb48aea0c..a5e62c28d2fa23a5ef9b9e2a0281b025aa943a30 100755
--- a/examples/tiny/s0/local/export.sh
+++ b/examples/tiny/s0/local/export.sh
@@ -13,13 +13,7 @@ ckpt_path_prefix=$2
 jit_model_export_path=$3
 model_type=$4
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/export.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/tiny/s0/local/test.sh b/examples/tiny/s0/local/test.sh
index b5b68c599c45ab50aa12ee35c120e02fb68740b4..4d00f30b852da5a370f5d4934f3caadd2b833c00 100755
--- a/examples/tiny/s0/local/test.sh
+++ b/examples/tiny/s0/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 model_type=$3
@@ -23,8 +19,7 @@ if [ $? -ne 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/test.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/tiny/s0/local/train.sh b/examples/tiny/s0/local/train.sh
index 9a76c7ade3b8d33d4c9828c29f941bdfc580814a..5b87780aebb1022e7914d6bd702faf7da7d21c14 100755
--- a/examples/tiny/s0/local/train.sh
+++ b/examples/tiny/s0/local/train.sh
@@ -10,17 +10,11 @@ source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 if [ ${seed} != 0  ]; then
     export FLAGS_cudnn_deterministic=True
     echo "using seed $seed & FLAGS_cudnn_deterministic=True ..."
 fi
 
-
 if [ $# != 3 ];then
     echo "usage: CUDA_VISIBLE_DEVICES=0 ${0} config_path ckpt_name model_type"
     exit -1
@@ -33,7 +27,6 @@ model_type=$3
 mkdir -p exp
 
 python3 -u ${BIN_DIR}/train.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/tiny/s1/local/align.sh b/examples/tiny/s1/local/align.sh
index ad6c84bc8cf398cddecfadbbb47b7ef9c60e9158..279461aafe19967ac054df815b08e60e351fcc7f 100755
--- a/examples/tiny/s1/local/align.sh
+++ b/examples/tiny/s1/local/align.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 
@@ -22,8 +18,7 @@ mkdir -p ${output_dir}
 # align dump in `result_file`
 # .tier, .TextGrid dump in `dir of result_file`
 python3 -u ${BIN_DIR}/alignment.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/tiny/s1/local/export.sh b/examples/tiny/s1/local/export.sh
index f99a15bade1c89f968e84a6c10d500466f884d5b..b562218e7a76c91d3b906d2c099218bf492627a8 100755
--- a/examples/tiny/s1/local/export.sh
+++ b/examples/tiny/s1/local/export.sh
@@ -12,13 +12,7 @@ config_path=$1
 ckpt_path_prefix=$2
 jit_model_export_path=$3
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 python3 -u ${BIN_DIR}/export.py \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --checkpoint_path ${ckpt_path_prefix} \
diff --git a/examples/tiny/s1/local/test.sh b/examples/tiny/s1/local/test.sh
index 4d3ed081a77e8bd93efa153818685537a66c6f11..19872bb36d4bfd0b6bcb20694a9f99715a591477 100755
--- a/examples/tiny/s1/local/test.sh
+++ b/examples/tiny/s1/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 
@@ -35,8 +31,7 @@ for type in attention ctc_greedy_search; do
         batch_size=64
     fi
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
@@ -52,8 +47,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     echo "decoding ${type}"
     batch_size=1
     python3 -u ${BIN_DIR}/test.py \
-    --device ${device} \
-    --nproc 1 \
+    --nproc ${ngpu} \
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/tiny/s1/local/train.sh b/examples/tiny/s1/local/train.sh
index 5097d4d030cab74fe5c3d52b223eddd1c27ea68d..71af3a006deec9032fcefd561fc46cea274d1e10 100755
--- a/examples/tiny/s1/local/train.sh
+++ b/examples/tiny/s1/local/train.sh
@@ -12,11 +12,6 @@ source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
-
 if [ ${seed} != 0  ]; then
     export FLAGS_cudnn_deterministic=True
     echo "using seed $seed & FLAGS_cudnn_deterministic=True ..."
@@ -34,7 +29,6 @@ mkdir -p exp
 
 python3 -u ${BIN_DIR}/train.py \
 --seed ${seed} \
---device ${device} \
 --nproc ${ngpu} \
 --config ${config_path} \
 --output exp/${ckpt_name} \
diff --git a/examples/v18_to_v2x/deepspeech2x/model.py b/examples/v18_to_v2x/deepspeech2x/model.py
index 1fe1e2d682644580f2e61ca66db64282923989cf..cbbc502d2e8e29479b091afb7a3a02fe9c2c85b1 100644
--- a/examples/v18_to_v2x/deepspeech2x/model.py
+++ b/examples/v18_to_v2x/deepspeech2x/model.py
@@ -401,7 +401,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
     def setup(self):
         """Setup the experiment.
         """
-        paddle.set_device(self.args.device)
+        paddle.set_device('gpu' if self.args.nprocs > 0 else 'cpu')
 
         self.setup_output_dir()
         self.setup_checkpointer()
diff --git a/examples/v18_to_v2x/exp_aishell/local/test.sh b/examples/v18_to_v2x/exp_aishell/local/test.sh
index 9fd0bc8d5bcdded1d33990a8ae20101d0a538441..2ae0740b3e8d44ab03e45f4c1b5dbb945657705e 100755
--- a/examples/v18_to_v2x/exp_aishell/local/test.sh
+++ b/examples/v18_to_v2x/exp_aishell/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 model_type=$3
@@ -23,8 +19,7 @@ if [ $? -ne 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/test.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/v18_to_v2x/exp_baidu_en8k/local/test.sh b/examples/v18_to_v2x/exp_baidu_en8k/local/test.sh
index b5b68c599c45ab50aa12ee35c120e02fb68740b4..4d00f30b852da5a370f5d4934f3caadd2b833c00 100755
--- a/examples/v18_to_v2x/exp_baidu_en8k/local/test.sh
+++ b/examples/v18_to_v2x/exp_baidu_en8k/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 model_type=$3
@@ -23,8 +19,7 @@ if [ $? -ne 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/test.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/examples/v18_to_v2x/exp_librispeech/local/test.sh b/examples/v18_to_v2x/exp_librispeech/local/test.sh
index b5b68c599c45ab50aa12ee35c120e02fb68740b4..4d00f30b852da5a370f5d4934f3caadd2b833c00 100755
--- a/examples/v18_to_v2x/exp_librispeech/local/test.sh
+++ b/examples/v18_to_v2x/exp_librispeech/local/test.sh
@@ -8,10 +8,6 @@ fi
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-device=gpu
-if [ ${ngpu} == 0 ];then
-    device=cpu
-fi
 config_path=$1
 ckpt_prefix=$2
 model_type=$3
@@ -23,8 +19,7 @@ if [ $? -ne 0 ]; then
 fi
 
 python3 -u ${BIN_DIR}/test.py \
---device ${device} \
---nproc 1 \
+--nproc ${ngpu} \
 --config ${config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
diff --git a/tests/chains/ds2_params_lite_train_infer.txt b/tests/chains/ds2_params_lite_train_infer.txt
index 82a9da9a98361ac868fc573e9230c0c7c1e01ca6..47f7db95b15f3a741b7b0ed56ca8ceb016258983 100644
--- a/tests/chains/ds2_params_lite_train_infer.txt
+++ b/tests/chains/ds2_params_lite_train_infer.txt
@@ -13,7 +13,7 @@ null:null
 null:null
 ##
 trainer:norm_train
-norm_train: ../../../deepspeech/exps/deepspeech2/bin/train.py --nproc 1 --config conf/deepspeech2.yaml --model_type offline --device gpu
+norm_train: ../../../deepspeech/exps/deepspeech2/bin/train.py --nproc 1 --config conf/deepspeech2.yaml --model_type offline
 pact_train:null
 fpgm_train:null
 distill_train:null
@@ -21,7 +21,7 @@ null:null
 null:null
 ##
 ===========================eval_params===========================
-eval: ../../../deepspeech/exps/deepspeech2/bin/test.py --nproc 1 --config conf/deepspeech2.yaml --result_file tests/9.rsl  --model_type offline --device gpu
+eval: ../../../deepspeech/exps/deepspeech2/bin/test.py --nproc 1 --config conf/deepspeech2.yaml --result_file tests/9.rsl  --model_type offline
 null:null
 ##
 ===========================infer_params===========================